Personalized Cine Cardiac MRI Protocol Optimization Using a Retrieval-Augmented Large Language Model: A Prospective Comparison With Technologist-Optimized Protocol.
Authors
Affiliations (3)
Affiliations (3)
- Radiation Sciences Department, School of Allied Medical Sciences, Tehran University of Medical Sciences, Tehran, Iran.
- Tehran Heart Center, Tehran University of Medical Sciences, Tehran, Iran.
- Ibn Al-Baladi for Children and Women Hospital, Baghdad, Iraq.
Abstract
Cine cardiac magnetic resonance imaging (MRI) protocols are typically based on static templates, which may not optimally address individual patient limitations such as arrhythmia or limited breath-hold capacity. To test whether an artificial intelligence (AI) framework using a large language model with retrieval-augmented generation (LLM-RAG) can provide cine cardiac MRI protocols that are equivalent or noninferior to those of technologist-optimized protocols. Prospective, intra-individual comparative, nonrandomized. Sixty-eight patients (mean age 46 ± 18 years; 39 women) referred for clinical cardiac MRI, including patients with normal function as well as those with arrhythmia, implants, or limited breath-hold capacity. 1.5 T, cine cardiac MRI. Each participant underwent cine cardiac MRI twice in one session. First with a technologist-optimized protocol, then with an AI-optimized protocol (fixed order). Image quality and artifact suppression were assessed using four-point Likert scales by two blinded readers. Left ventricular (LV) functional parameters and repeat acquisitions were recorded. Wilcoxon signed-rank test, McNemar's test, paired t-tests, and Bland-Altman analysis. Significance set at p < 0.05. Compared to technologist-, AI-optimized cine protocols demonstrated significantly higher image quality (3.69 ± 0.47 vs. 3.00 ± 0.49, p < 0.001) and artifact suppression (3.19 ± 0.35 vs. 2.17 ± 0.45, p < 0.001). The AI-optimized protocol reduced repeat acquisitions by 68% (6 vs. 19 repeats, p = 0.028). No significant differences were observed in LV functional parameters between both protocols (all p > 0.05). Bland-Altman analysis showed minimal bias for all LV functional parameters. In this nonrandomized study, a LLM-RAG framework generated cine cardiac MRI protocols that were noninferior to technologist-optimized protocols for image quality and demonstrated high agreement for LV functional measurements with fewer repeat acquisitions. These findings demonstrated the feasibility of AI-guided patient-adaptive protocoling. Randomized trials are warranted. 2. Stage 2.